Three hundred and eighteen million metric tons of carbon dioxide a year. That is the combined output of the gas-fired power plants now being proposed to feed America’s AI data centers, according to figures from BloombergNEF reported by Fortune this week. The research firm is tracking 99 such plants. Built and run as planned, they would lift US power sector emissions by roughly 20 percent. Run flat out, the increase approaches a third.
For an industry that spent a decade advertising its wind and solar contracts, that is a remarkable turn. Before the AI boom, cloud companies deliberately sited data centers near renewable power and competed on sustainability pledges. The AI buildout has broken that pattern. Amazon and Microsoft are now among the developers turning to bespoke natural gas plants, built specifically to serve individual data center campuses.
Why gas, and why now
The short answer is speed. Connecting a large new load to the grid can take years of waiting in interconnection queues, and AI developers are not in a waiting mood. A dedicated gas plant behind the fence skips the queue entirely. It also runs around the clock, which matters for training clusters that companies want lit at all hours to justify their cost.
The scale of the buildout keeps growing regardless of what powers it. This week alone, Nvidia backed OpenAI’s Ohio data center expansion with a $105 billion guarantee and Micron announced a $10 billion investment in US memory research. Globally, data centers consumed about 1.5 percent of electricity in 2024. Estimates for the United States suggest data centers could account for somewhere between 9.5 and 15.3 percent of national electricity use by 2030. The spread in those numbers is itself telling: nobody, including the utilities signing the contracts, knows exactly how much power AI will actually need.
The bigger number nobody prices in
Even 318 million tons may understate AI’s climate exposure. A peer-reviewed study published this month in npj Climate Action modeled something the data center debate mostly ignores: what AI does to the industries that use it. The researchers treated AI as a productivity amplifier for both fossil fuel extraction and renewable energy, then ran the numbers through a global energy-economy model.
The result cuts against the industry’s favorite talking point. Emissions enabled by AI making fossil fuel extraction faster and cheaper came out 3.3 to 13.3 times larger than the emissions from AI data centers themselves. If AI is adopted at similar rates across both sectors, net global emissions rise by 0.47 to 1.8 gigatonnes of CO2 per year, between 1.2 and 4.8 percent of global energy-related emissions. For AI to cut emissions on net, the study found, its productivity gains for renewables would need to be four to five times greater than its gains for the fossil sector.
Put differently: the argument that AI will eventually invent its way out of its own carbon footprint requires the technology to help clean energy far more than it helps oil and gas. So far there is no evidence the balance tilts that way.
The backlash arrives on schedule
The politics are catching up. A bill introduced in the House this week would impose a federal electricity tax on data centers, and Axios reports that local fights over siting, water and power costs have reached fever pitch in communities across the country. Ratepayers are noticing that new gas capacity built for a private customer still shapes the grid everyone shares.
The industry’s counterarguments exist, and some are serious. Flexible data center loads that power down during grid stress could ease the crunch. Nuclear deals keep multiplying, including TerraPower’s bet on powering AI campuses with advanced reactors. But reactors take a decade to build and flexibility remains mostly a white paper. Gas turbines ship now.
The number to watch is not 99. It is how many of those 99 plants actually clear permitting, financing and turbine supply constraints over the next two years. Every one that does locks in decades of emissions for compute that was supposed to help solve the problem.
Supply chains may end up doing what regulation has not. Gas turbine manufacturers are sold out years in advance, with order books stretching toward the end of the decade, and prices for new units have climbed accordingly. Several of the 99 proposed plants are competing for the same limited production slots, which means some of the emissions in BloombergNEF’s tally may never materialize simply because the hardware cannot be delivered on the timelines AI developers demand.
For more coverage of AI and climate, visit Mylistingo.
